A log-adjusted t-statistic for large clinical laboratory datasets: a simulation study and real-world application.

Area of Science:

• Biostatistics
• Clinical Laboratory Science
• Statistical Modeling

Background:

• Conventional t-tests in large datasets often produce statistically significant results that are practically trivial due to increased statistical power with sample size.
• This oversensitivity can obscure the detection of substantively meaningful effects when sample sizes are very large.

Purpose of the Study:

• To evaluate a log-adjusted t-statistic as a sample-size-aware modification of the classical t-statistic.
• To assess its performance in reducing oversensitivity to sample size in large datasets and its utility in clinical laboratory data interpretation.

Main Methods:

• Monte Carlo simulations were conducted for two-sample comparisons across a wide range of sample sizes (10 to 50,000) and effect sizes (δ = 0 to 1.0).
• The performance of the log-adjusted t-statistic was compared against the classical t-test.
• The method was applied to a real clinical laboratory dataset with 464,145 participants.

Main Results:

• The log-adjusted t-statistic demonstrated null rejection rates close to the nominal 0.05 level across all sample sizes, unlike the classical t-test which became oversized at very large sample sizes.
• The adjustment was more conservative for small effects but maintained high rejection rates for larger effects.
• In real-data analysis, several highly significant sex differences with small effect sizes (e.g., platelet count, potassium) yielded reference p-values above 0.05 after adjustment, while larger effects (e.g., hematocrit, HDL cholesterol) remained significant.

Conclusions:

• The log-adjusted t-statistic effectively attenuates sample-size-driven significance in large datasets.
• It preserves the detection of substantively meaningful effects, serving as a valuable empirical decision aid for interpreting clinical laboratory data.
• This approach helps distinguish true biological variation from statistically significant but practically irrelevant findings.